AICodingAgents

Global app state. Contains configured MCP servers, per-project allowed tools and trust decisions, list of every project directory Claude Code has been run in, OAuth account metadata (email/org, no tokens), install method, and usage counters. Backup copies match .claude.json.backup*

One subfolder per project (path with separators replaced by dashes). Each session is <sessionId>.jsonl containing every user prompt, assistant response, tool call and tool result. Subfolders per session hold subagent transcripts (subagents\agent-*.jsonl), oversized tool outputs (tool-results\), session-memory summaries and the per-project auto memory (memory\*.md)

Global index of every prompt typed into Claude Code with timestamp, project path, and pasted content references

settings.json and settings.local.json: permission allow/deny lists, hooks, environment variables, enabled plugins, model and telemetry configuration

User-level instructions loaded into every session

Custom keyboard shortcuts

Aggregated usage statistics (sessions, messages, tokens, models used per day)

User-defined scripts executed automatically on tool use, prompt submit, session start/stop, etc. Review for persistence or data exfiltration

User-defined slash commands (markdown prompts)

User-installed skills (SKILL.md plus supporting scripts). Can be large if many skills are installed

User-defined subagent definitions and their persistent agent-memory

User-level rule files loaded into every project

User-level dynamic workflow scripts that orchestrate subagents

installed_plugins.json, known_marketplaces.json, blocklist.json, config.json and plugin-catalog-cache.json. Records which third-party plugins and marketplaces were installed and from where. Plugin source trees under plugins\cache and plugins\repos are intentionally not collected

Snapshot of the user's shell environment (functions, aliases, PATH) captured at session start. One file per session

Pre-edit copies of files modified by Claude Code, organized by session ID with versioned entries (<hash>@v1, @v2...). Allows reconstruction of what the agent changed

Per-session task lists (older versions)

Per-session task state, background task output and lock files (newer versions)

Plan-mode documents written before implementation

Live session registry keyed by PID: session ID, working directory, start time, and connected IDE. Per-session .key files are not collected

IDE integration lock files, one per IDE extension instance, containing the IDE PID, workspace folders and auth token used for the local IDE bridge

Per-session debug logs when present

Content pasted into the prompt that was too large to inline in history.jsonl

Per-session environment variable overrides

Automatic backups of settings files, named by epoch millisecond timestamp

Feature-flag and telemetry cache containing stable device ID and last session metadata

Global storage for the Claude Code VS Code extension (publisher.name anthropic.claude-code)

Global storage for the Claude Code extension when installed in Cursor

Claude Code sessions launched from the Claude Desktop app (Code tab). Organized by account and session ID

Claude Desktop Cowork (local agent mode) session state. One JSON per turn plus spaces.json. Cowork runs the Claude Code agent against local files

MCP servers configured for Claude Desktop, which are also exposed to embedded Claude Code and Cowork sessions

Same as above for Microsoft Store / MSIX installs of Claude Desktop, where AppData\Roaming is virtualized under the package LocalCache

Same as above for Microsoft Store / MSIX installs of Claude Desktop

Same as above for Microsoft Store / MSIX installs of Claude Desktop